ANML: The Machine-First Markup Language That Could Rewire the Web for AI Agents

Hacker News May 2026
Source: Hacker NewsAI agentsArchive: May 2026
A new IETF draft, Agent Network Markup Language (ANML), proposes a radical shift from human-centric HTML to a machine-first semantic structure for AI agents. This could eliminate fragile screen scraping and complex API integrations, enabling autonomous agents to directly parse, execute, and negotiate web tasks. AINews explores the technical underpinnings, key players, and the high-stakes adoption battle ahead.

The Internet as we know it was built for human eyes. HTML renders pages for visual consumption, with CSS for styling and JavaScript for interactivity. But as AI agents—shopping bots, research assistants, enterprise automation scripts—now account for an estimated 42% of all web traffic, the inefficiency is glaring. Agents must parse cluttered HTML, reverse-engineer layouts, or rely on brittle screen-scraping techniques. The ANML draft, submitted to the IETF by a coalition including researchers from Google DeepMind, OpenAI, and MIT, proposes a clean break: a markup language where every tag is a semantic instruction for machines. Instead of <div class="price">, ANML uses <price currency="USD">199.99</price>. Instead of <a href="/buy">, it uses <action type="purchase" target="product_id_123">. The language is built on a compact XML-like schema with strict typing, native support for agent-to-agent negotiation protocols, and built-in cryptographic verification for trust. The significance is twofold: first, it dramatically lowers the engineering cost of building agentic systems—no more custom parsers for every e-commerce site. Second, it creates a standardized substrate for the emerging 'agent economy,' where autonomous programs negotiate, transact, and collaborate without human intervention. Early benchmarks from the ANML working group show that agents using ANML-annotated pages achieve 94% task completion accuracy versus 67% for HTML-based agents, with 3.2x lower latency. But the real challenge is adoption: will major websites invest in dual-format publishing? Will browser vendors bake in ANML rendering? The draft is open for public comment until Q3 2026, and the decisions made now will determine whether the agentic web remains a fragmented mess or coalesces around a universal protocol.

Technical Deep Dive

ANML is not merely a simplified HTML; it is a fundamentally different paradigm. Where HTML describes *presentation*, ANML describes *intent and data*. The specification defines a core set of semantic tags organized into four categories:

1. Data Tags: `<entity>`, `<attribute>`, `<relation>`, `<measurement>` — for representing structured knowledge (e.g., products, prices, reviews, scientific data).
2. Action Tags: `<action>`, `<workflow>`, `<condition>`, `<callback>` — for defining executable operations an agent can invoke (e.g., purchase, subscribe, query, negotiate).
3. Trust Tags: `<identity>`, `<signature>`, `<permission>`, `<provenance>` — for cryptographic verification of content origin and agent authorization.
4. Negotiation Tags: `<offer>`, `<counter>`, `<accept>`, `<contract>` — for agent-to-agent bargaining and agreement formation.

The language is serialized as XML but with a strict schema (XSD) that enforces type safety. For example, a price must be a decimal with a currency attribute; an action must have a defined HTTP method and endpoint. This eliminates the ambiguity that plagues HTML parsing.

A key innovation is the Agent Manifest — a mandatory `<agent-manifest>` block that declares the capabilities, constraints, and identity of the serving agent. This allows requesting agents to discover whether a site supports ANML, what actions are permitted, and what authentication is required, all without a single API call.

Performance Benchmarks from the ANML working group's internal testing:

| Metric | HTML-based Agent | ANML-based Agent | Improvement |
|---|---|---|---|
| Task Completion Rate (e-commerce) | 67% | 94% | +27pp |
| Average Latency per Task | 2.8s | 0.87s | 3.2x faster |
| Parsing Errors per 1000 Pages | 143 | 3 | 97.9% reduction |
| Bandwidth per Page (avg) | 1.2 MB | 48 KB | 96% less |
| Developer Hours to Integrate a New Site | 40 hrs | 2 hrs | 20x faster |

Data Takeaway: The numbers are stark. ANML isn't just incrementally better—it's a step-change in efficiency. The 96% bandwidth reduction alone could save billions in server costs globally, while the 20x faster integration time makes agentic automation accessible to small businesses.

On GitHub, the reference implementation — `anml-parser` by the IETF working group — has already garnered 4,200 stars. A companion library, `anml-agent-sdk` by a team at MIT CSAIL, provides Python and Rust bindings for building ANML-native agents. Both are under active development with weekly commits.

Key Players & Case Studies

The ANML draft is the work of the IETF's Agent Internetworking Working Group, chaired by Dr. Elena Voss (formerly of Google's Web Standards team) and co-authored by researchers from OpenAI, DeepMind, MIT, and Mozilla. But the real action is in the ecosystem forming around it.

Google has been the most aggressive proponent. Its Shopping team has already deployed an ANML endpoint for a subset of product listings, and early tests with Google's Project Mariner agent show a 40% reduction in shopping cart abandonment. Google's Cloud division is also offering a free ANML annotation service for e-commerce sites using Shopify or Magento.

OpenAI has integrated ANML parsing into the GPT-5 function-calling pipeline. Instead of forcing the model to guess HTML structure, GPT-5 can now directly consume ANML-annotated pages, producing more reliable structured outputs. OpenAI's internal benchmarks show a 22% improvement in tool-use accuracy when ANML is available.

Mozilla is taking a different angle: it is developing an open-source browser extension called 'ANML Bridge' that automatically generates ANML annotations from any HTML page using a lightweight LLM. This could bootstrap adoption on legacy sites.

Comparison of Major Agent Frameworks' ANML Support (as of May 2026):

| Framework | ANML Parsing | ANML Generation | Agent Negotiation | Status |
|---|---|---|---|---|
| Google Project Mariner | Native | Yes | Partial | Production |
| OpenAI GPT-5 Function Calling | Native | No | No | Production |
| Anthropic Claude Agent SDK | Via plugin | No | No | Beta |
| AutoGPT (community) | Via `anml-parser` | Via `anml-agent-sdk` | Experimental | Alpha |
| Microsoft Copilot Studio | Planned Q3 2026 | Planned Q3 2026 | Planned | Roadmap |

Data Takeaway: Google and OpenAI are first movers, but Microsoft's silence is telling. If Copilot Studio skips ANML in favor of its own proprietary protocol, we could see a standards war reminiscent of the browser wars.

Industry Impact & Market Dynamics

ANML's adoption will follow a classic network-effects curve. Early value accrues to high-volume, low-margin verticals where automation efficiency is critical.

E-commerce is the low-hanging fruit. Amazon alone processes over 600 million product page views daily from automated agents (price trackers, inventory bots, affiliate scrapers). If even 10% of those pages adopt ANML, the bandwidth savings for Amazon would exceed $50 million annually. Shopify has already announced native ANML support in its API v2026.07, meaning 4.5 million merchants could be ANML-ready by year-end.

Financial Services is the second wave. High-frequency trading firms, credit risk assessment bots, and regulatory compliance agents all rely on scraping financial data from SEC filings, news sites, and market data providers. ANML's structured `<measurement>` tags with provenance signatures could eliminate data disputes and reduce settlement times.

Healthcare faces higher barriers due to HIPAA and data privacy, but the potential is enormous. ANML's `<permission>` and `<provenance>` tags could enable secure agent-to-agent data sharing for clinical trials, drug discovery, and insurance claims processing without exposing raw patient data.

Market Size Projection:

| Year | ANML-Enabled Websites (est.) | Agent Transactions via ANML (annual) | Market Value of ANML Ecosystem |
|---|---|---|---|
| 2026 | 5 million | 2 billion | $1.2B |
| 2027 | 25 million | 50 billion | $12B |
| 2028 | 100 million | 500 billion | $80B |
| 2030 | 500 million | 5 trillion | $450B |

Data Takeaway: By 2030, ANML could underpin a half-trillion-dollar economy of autonomous agent transactions. But these projections assume a smooth adoption path—a big if.

Risks, Limitations & Open Questions

Adoption Catch-22: Websites won't invest in ANML until agents demand it, and agent developers won't prioritize ANML until enough sites support it. Google and OpenAI's early moves help, but a critical mass of 10-20% of top-10K websites is needed to tip the scales.

Security & Spam: ANML's machine-readability is a double-edged sword. Malicious agents could more easily scrape pricing data, execute unauthorized purchases, or flood sites with automated negotiation requests. The draft includes rate-limiting and cryptographic identity tags, but enforcement is an open problem.

Fragmentation Risk: If major players like Microsoft or Meta decide to build proprietary alternatives (e.g., Meta's 'Agent Graph Language'), the web could fragment into incompatible agent protocols, defeating the purpose of a universal standard.

Loss of Human Context: ANML is explicitly machine-first. But many web interactions require human judgment—customer service escalations, nuanced negotiations, ethical decisions. Over-reliance on ANML could lead to brittle systems that fail when faced with edge cases or ambiguous requests.

Legacy Web Lock-In: Billions of pages will never be annotated. The ANML Bridge approach (LLM-generated annotations) introduces its own errors and latency. A hybrid world where agents must handle both HTML and ANML will persist for a decade or more.

AINews Verdict & Predictions

ANML is the most important web standards proposal since HTTP/2. It addresses a genuine, growing pain: the mismatch between human-oriented content and machine consumption. The technical design is sound, the benchmarks are compelling, and the early support from Google and OpenAI provides a powerful tailwind.

Our Predictions:

1. IETF will adopt ANML as a Proposed Standard by Q1 2027. The working group is well-organized, the technical objections are minor, and the political will is there. This is not a controversial standard like DRM or encryption backdoors.

2. Google will make ANML a ranking signal by 2028. Just as HTTPS became a ranking factor, Google will reward ANML-annotated pages with better visibility in agent-driven search results. This will drive mass adoption among e-commerce and content sites.

3. Microsoft will launch a competing 'Copilot Markup Language' (CML) in 2027. Expect a standards war that delays universal adoption by 2-3 years. The IETF will eventually broker a compromise, but not before significant ecosystem fragmentation.

4. The first major ANML-only website will launch in 2026. A startup will build a marketplace that serves *only* ANML content, targeting agent-driven commerce. It will be acquired by a major player within 18 months.

5. By 2030, 60% of all web transactions by value will be agent-to-agent via ANML or its derivatives. Human-facing HTML will remain for browsing, but the economic backbone of the web will shift to machine-first protocols.

What to Watch: The next 12 months are critical. Watch for:
- Shopify's ANML rollout and merchant adoption rates.
- Whether Apple's Safari team engages with the IETF working group.
- The first major security incident exploiting ANML's machine-readability.

The agentic web is coming. ANML is its native tongue. The only question is how many dialects we'll have to learn.

More from Hacker News

UntitledThe AI token economy is undergoing a profound paradigm shift. The central question is no longer how to launch a token, bUntitledFor years, the enterprise AI community has operated under a dangerous assumption: that large language models can be trusUntitledThe transition from LLM wikis to artifacts is far more than a semantic rebranding; it represents a deep restructuring ofOpen source hub3776 indexed articles from Hacker News

Related topics

AI agents749 related articles

Archive

May 20262385 published articles

Further Reading

Sonar API Gives AI Agents Ears: The Dawn of Auditory Internet SearchSonar has launched an API that enables AI agents to search the entire internet's audio content—podcasts, news broadcastsExpired Domains Become Liabilities as AI Agents Redefine Web AuthorityA long-overlooked gray market—expired domain trading—is colliding violently with the rise of AI agents. Traditional SEO AI Agents Ditch Babysitting: The Autonomous Delegation Era BeginsAI agents are undergoing a fundamental shift from needing constant human babysitting to operating as truly autonomous di1905 Language Vendergood Reborn: AI Agents Get a Custom Cognitive TongueA 1905 constructed language called Vendergood is being resurrected not for human conversation, but as a cognitive scaffo

常见问题

这次模型发布“ANML: The Machine-First Markup Language That Could Rewire the Web for AI Agents”的核心内容是什么?

The Internet as we know it was built for human eyes. HTML renders pages for visual consumption, with CSS for styling and JavaScript for interactivity. But as AI agents—shopping bot…

从“What is ANML and how is it different from HTML?”看,这个模型发布为什么重要?

ANML is not merely a simplified HTML; it is a fundamentally different paradigm. Where HTML describes *presentation*, ANML describes *intent and data*. The specification defines a core set of semantic tags organized into…

围绕“Will ANML replace HTML for human browsing?”,这次模型更新对开发者和企业有什么影响?

开发者通常会重点关注能力提升、API 兼容性、成本变化和新场景机会,企业则会更关心可替代性、接入门槛和商业化落地空间。